How otrf/genai-security-adventures works
Main Components
• Jupyter Book & Documentation: The root contains configuration files (_config.yml, _toc.yml) and a README that indicates this is a Jupyter Book project serving as a documentation and presentation hub.
• Experiments: This is the heart of the repository. Under the “experiments” directory, you see distinct subdirectories:
• Agents (with further experiments like Chain-of-Thought and tool-calling agents)
• Basics (containing several introductory experiments and supporting scripts)
• Bloodhound-GPT (which includes database creation logic and a Dockerfile, indicating heavier integration and possibly stateful components)
• ExternalAPIs (demonstrating integrations, e.g., with MSGraph, with supporting OpenAPI specs)
• Local-GPT (a collection of experiments running local variants of generative models)
• RAG (focused on Threat Intelligence with Retrieval-Augmented Generation experiments, including extensive document sets and notebooks)
• ROADTools-GPT (which has database components, suggesting its role in network or reconnaissance tasks)
• Getting Started: A lightweight guide that introduces fundamental concepts in AI and deep learning, which supports the hands-on experiments.
• Images: Supporting assets for documentation or instructions (API keys illustration).
• Document the high-level separation between the documentation (Jupyter Book) and the experiments. Represent the Jupyter Book as a central hub that references all experiments.
• Agents Module: Highlighting the different strategic approaches (CoT, ReAct, Tool-Calling) and the notebooks that demonstrate each.
• Basics Module: Represent experiments that illustrate fundamental interactions with LLMs (chat completions, sentiment analysis, tokenizing).
• Threat Intelligence/RAG Module: Show a detailed view of how document sources, intermediate processing (source-knowledge, templates), and notebook-based analyses interact.
• Local Integration Components: Sketch how local model experiments (Local-GPT, Ollama setups) are separated from cloud-based or API-driven experiments.
• Database/Stateful Components: In Bloodhound-GPT and ROADTools-GPT, emphasize the handling of databases and associated scripts/tools (like DBCreator in Bloodhound-GPT and the roadrecon.db in ROADTools-GPT).
• Cross-Cutting Concerns:
• Security and LLM integration as a conceptual overlay across all experiments.
• Reusable code scripts (e.g., m365d.py, requirements.txt) that ensure consistency for local dependencies.